Check out the new USENIX Web site.

AFRAID - A Frequently Redundant Array of Independent Disks

Stefan Savage
University of Washington, Seattle, WA

John Wilkes
Hewlett-Packard Laboratories, Palo Alto, CA

Abstract

Disk arrays are commonly designed to ensure that stored data will always be able to withstand a disk failure, but meeting this goal comes at a significant cost in performance. We show that this is unnecessary. By trading away a fraction of the enormous reliability provided by disk arrays, it is possible to achieve performance that is almost as good as a non-parity-protected set of disks.

In particular, our AFRAID design eliminates the small-update penalty that plagues traditional RAID 5 disk arrays. It does this by applying the data update immediately, but delaying the parity update to the next quiet period between bursts of client activity. That is, AFRAID makes sure that the array is frequently redundant, even if it isn't always so. By regulating the parity update policy, AFRAID allows a smooth trade-off between performance and availability.

Under real-life workloads, the AFRAID design can provide close to the full performance of an array of unprotected disks, and data availability comparable to a traditional RAID 5. Our results show that AFRAID offers 42% better performance for only 10% less availability, 97% better for 23% less, and as much as a factor of 4.1 times better performance for giving up less than half RAID 5's availability.

We explore here the detailed availability and performance implications of the AFRAID approach.

Download the full text of this paper in PDF, ASCII (57,221 bytes) and POSTSCRIPT (340,685 bytes) form.

To Become a USENIX Member, please see our Membership Information.